Methods and Systems for Varnish Analysis of Stator Images

1. A method for automatically analyzing images of a stator, comprising: receiving images of the stator at a processor of a computing system, the images depicting cross-sections of the stator; processing the images using deep learning algorithms by cropping and filtering a region of the images corresponding to slots of the stator and converting the images to one or more of cluster-only images and binary masks; feeding the one or more of the cluster-only images and binary masks to an artificial intelligence (AI) model implemented at the processor; obtaining one or more of varnish estimates and void estimates from the AI model to generate a training dataset; and training a deep learning tool, based on the training dataset, to estimate varnish fill percentages from the images and display the varnish fill percentages at a display device.

2. The method of claim 1, wherein the cross-sections are one or more of transverse cross-sections, obtained by slicing the stator along one or more planes perpendicular to a central axis of rotation of the stator, and axial cross-sections, obtained by slicing the stator through at least one slot of the stator along the central axis of rotation of the stator.

3. The method of claim 2, wherein the transverse cross-sections depict cross-sectional areas of conductors of the stator, the conductors surrounded by insulating paper, and wherein the axial cross-sections depict a surface of the at least one slot.

4. The method of claim 1, wherein converting the images to the cluster-only images includes applying k-means clustering to the images, and wherein applying k-means clustering includes converting the images to hue saturation value (HSV) color space and identifying clusters in the images.

5. The method of claim 4, wherein a cluster-only image is generated for each of the clusters, and wherein feeding the cluster-only images to the AI model includes analyzing the cluster-only images for each of the clusters according to a plurality of image parameters and inputting analysis results from the analyzing to the AI model.

6. The method of claim 5, wherein obtaining the void estimates from the AI model includes obtaining a cluster-only image corresponding to varnish as an output from the AI model and filling in gaps at a paper border of the output to create a continuous border at the paper border.

7. The method of claim 6, further comprising generating a paper fill mask by applying the continuous border to a binary image, the binary image formed from the output from the AI model, to create a gap image, and wherein the gap image is combined with an original image of the images of the stator to generate the paper fill mask.

8. The method of claim 7, wherein the paper fill mask includes voids, and wherein obtaining the void estimates includes assessing and quantifying properties of the voids.

9. The method of claim 1, wherein converting the images to binary masks includes converting the images to red, green, blue (RGB) color space to generate a differential image and analyzing the differential image based on a plurality of image parameters.

10. The method of claim 9, wherein obtaining the varnish estimates from the AI model includes inputting analysis results of the analyzing of the differential image to the AI model to obtain a threshold value, and wherein the threshold value defines a pixel color boundary that differentiates between pixels corresponding to varnish and pixels not corresponding to varnish.

11. A system for evaluating a varnish condition of a stator, comprising: a housing enclosing a UV light source and digital imaging equipment; and a processor configured with executable instructions stored in non-transitory memory that, when executed, cause the processor to: receive images of cross-sections of the stator from the digital imaging equipment; process the images using deep learning algorithms by cropping and filtering a region of the images displaying at least one slot and converting the images to converted images, the converted images being one or more of cluster-only images and binary masks; input the converted images to a machine learning model trained to identify varnish in the converted images based on one or more of color distribution analysis and cluster analysis of the converted images; output one or more of a varnish estimate and a void estimate from the machine learning model to generate a training dataset for a deep learning model; and train the deep learning model based on the training dataset to train the deep learning model to estimate varnish fill percentages based on the images and display the varnish fill percentages in a report at a display device.

12. The system of claim 11, wherein two training techniques are used to train the machine learning model to output the varnish estimate when the cross-sections are axial cross-sections of the stator, the axial cross-sections obtained by slicing the stator along a central axis of rotation of the stator, and wherein a first training technique of the two training techniques includes generating the cluster-only images, the cluster-only images depicting clusters, plotting blue values for the clusters to determine a peak location and a peak distribution of each of the clusters, and determining a total number of peaks of the cluster-only images.

13. The system of claim 12, wherein the machine learning model is trained to assign a value within an inclusive range of 0 to 10 to each of the clusters, and wherein an assigned value of 2 or less indicates that a corresponding cluster represents varnish.

14. The system of claim 12, wherein in a second training technique of the two training techniques, differential images are obtained from the converted images by plotting the cluster-only images in red, green, blue (RGB) color space showing a contrast between red and green color spaces, and wherein the differential images are analyzed based on a blue portion of the differential images and analysis results are input to the machine learning model.

15. The system of claim 14, wherein the machine learning model is trained to determine a threshold value of a contrast between red and green, the threshold value being a value between and inclusive of 0 to 20, by applying random forests to data of the differential images, and wherein the machine learning model is further trained to apply the threshold value to the differential images to generate thresholded binary images, and wherein remaining non-white pixels in the thresholded binary images correspond to varnish.

16. The system of claim 12, wherein the machine learning model is trained according to one of the two training techniques based on a comparison of results from the two training techniques applied to a common fluorescence image, and wherein the comparison of results includes determining a total area of blobs in each of a cluster-only image, the cluster-only image generated via a first training technique of the two training techniques, and a differential image, the differential image generated via a second training technique of the two training techniques.

17. The system of claim 16, wherein the comparison of results further includes computing an extent and an average area of the blobs in each of the cluster-only image and the differential image, removing any blobs smaller than a threshold number of pixels from each of the cluster-only image and the differential image, and comparing a change in each of the cluster-only image and the differential image, and wherein one of the first training technique and the second training technique is selected that corresponds to which of the cluster-only image and the differential image exhibits a least amount of change.

18. The system of claim 11, wherein, when the cross-sections are transverse cross-sections of the stator, obtained by slicing the stator perpendicular to a central axis of rotation of the stator, the machine learning model is trained using images converted to red, green, blue (RGB) color space and using image data input to the machine learning model, the image data including one or more of a magnification of the images, a number of windings per slot of the stator, and a geometry of the windings.

19. A method for evaluating a varnish condition of a stator, comprising: illuminating a cross-section of the stator with light from a UV light source; obtaining a fluorescence image of the cross-section via digital imaging equipment and transmitting the fluorescence image to a processor; processing the fluorescence image, at the processor, by cropping the fluorescence image to borders corresponding to one or more slots identified in the fluorescence image; converting the processed image using one or more of color distribution analysis and clustering analysis to create a converted image; inputting data from the converted image to a machine learning model implemented at the processor, the machine learning model trained to locate and quantify one or more of varnish and voids based on the converted image; generating a training dataset based on an output from the machine learning model and feeding the training dataset to a deep learning tool to train the deep learning tool to estimate a varnish fill percentage from the fluorescence image; and obtaining the varnish fill percentage from the deep learning tool and displaying the varnish fill percentage at a display device as a visual representation.

20. The method of claim 19, wherein the machine learning model is trained, when the cross-section is a transverse cross-section of the stator, by obtaining one or more copper estimates via comparison of a known quantity of conductors the converted image to an estimated quantity of conductors, applying an error correction and smoothing to the converted image and quantifying voids in the converted image, and wherein the converted image is a binary mask.